xref: /netbsd-src/sys/uvm/uvm_pager.c (revision 23c8222edbfb0f0932d88a8351d3a0cf817dfb9e)
1 /*	$NetBSD: uvm_pager.c,v 1.64 2004/10/03 07:59:02 enami Exp $	*/
2 
3 /*
4  *
5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *      This product includes software developed by Charles D. Cranor and
19  *      Washington University.
20  * 4. The name of the author may not be used to endorse or promote products
21  *    derived from this software without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
24  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
25  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
26  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
27  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
28  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
32  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33  *
34  * from: Id: uvm_pager.c,v 1.1.2.23 1998/02/02 20:38:06 chuck Exp
35  */
36 
37 /*
38  * uvm_pager.c: generic functions used to assist the pagers.
39  */
40 
41 #include <sys/cdefs.h>
42 __KERNEL_RCSID(0, "$NetBSD: uvm_pager.c,v 1.64 2004/10/03 07:59:02 enami Exp $");
43 
44 #include "opt_uvmhist.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/proc.h>
49 #include <sys/malloc.h>
50 #include <sys/pool.h>
51 #include <sys/vnode.h>
52 
53 #define UVM_PAGER
54 #include <uvm/uvm.h>
55 
56 struct pool *uvm_aiobuf_pool;
57 
58 /*
59  * list of uvm pagers in the system
60  */
61 
62 struct uvm_pagerops * const uvmpagerops[] = {
63 	&aobj_pager,
64 	&uvm_deviceops,
65 	&uvm_vnodeops,
66 	&ubc_pager,
67 };
68 
69 /*
70  * the pager map: provides KVA for I/O
71  */
72 
73 struct vm_map *pager_map;		/* XXX */
74 struct simplelock pager_map_wanted_lock;
75 boolean_t pager_map_wanted;	/* locked by pager map */
76 static vaddr_t emergva;
77 static boolean_t emerginuse;
78 
79 /*
80  * uvm_pager_init: init pagers (at boot time)
81  */
82 
83 void
84 uvm_pager_init()
85 {
86 	u_int lcv;
87 	vaddr_t sva, eva;
88 
89 	/*
90 	 * init pager map
91 	 */
92 
93 	sva = 0;
94 	pager_map = uvm_km_suballoc(kernel_map, &sva, &eva, PAGER_MAP_SIZE, 0,
95 	    FALSE, NULL);
96 	simple_lock_init(&pager_map_wanted_lock);
97 	pager_map_wanted = FALSE;
98 	emergva = uvm_km_valloc(kernel_map, round_page(MAXPHYS));
99 	emerginuse = FALSE;
100 
101 	/*
102 	 * init ASYNC I/O queue
103 	 */
104 
105 	TAILQ_INIT(&uvm.aio_done);
106 
107 	/*
108 	 * call pager init functions
109 	 */
110 	for (lcv = 0 ; lcv < sizeof(uvmpagerops)/sizeof(struct uvm_pagerops *);
111 	    lcv++) {
112 		if (uvmpagerops[lcv]->pgo_init)
113 			uvmpagerops[lcv]->pgo_init();
114 	}
115 }
116 
117 /*
118  * uvm_pagermapin: map pages into KVA (pager_map) for I/O that needs mappings
119  *
120  * we basically just map in a blank map entry to reserve the space in the
121  * map and then use pmap_enter() to put the mappings in by hand.
122  */
123 
124 vaddr_t
125 uvm_pagermapin(pps, npages, flags)
126 	struct vm_page **pps;
127 	int npages;
128 	int flags;
129 {
130 	vsize_t size;
131 	vaddr_t kva;
132 	vaddr_t cva;
133 	struct vm_page *pp;
134 	vm_prot_t prot;
135 	UVMHIST_FUNC("uvm_pagermapin"); UVMHIST_CALLED(maphist);
136 
137 	UVMHIST_LOG(maphist,"(pps=0x%x, npages=%d)", pps, npages,0,0);
138 
139 	/*
140 	 * compute protection.  outgoing I/O only needs read
141 	 * access to the page, whereas incoming needs read/write.
142 	 */
143 
144 	prot = VM_PROT_READ;
145 	if (flags & UVMPAGER_MAPIN_READ)
146 		prot |= VM_PROT_WRITE;
147 
148 ReStart:
149 	size = npages << PAGE_SHIFT;
150 	kva = 0;			/* let system choose VA */
151 
152 	if (uvm_map(pager_map, &kva, size, NULL,
153 	      UVM_UNKNOWN_OFFSET, 0, UVM_FLAG_NOMERGE) != 0) {
154 		if (curproc == uvm.pagedaemon_proc) {
155 			simple_lock(&pager_map_wanted_lock);
156 			if (emerginuse) {
157 				UVM_UNLOCK_AND_WAIT(&emergva,
158 				    &pager_map_wanted_lock, FALSE,
159 				    "emergva", 0);
160 				goto ReStart;
161 			}
162 			emerginuse = TRUE;
163 			simple_unlock(&pager_map_wanted_lock);
164 			kva = emergva;
165 			/* The shift implicitly truncates to PAGE_SIZE */
166 			KASSERT(npages <= (MAXPHYS >> PAGE_SHIFT));
167 			goto enter;
168 		}
169 		if ((flags & UVMPAGER_MAPIN_WAITOK) == 0) {
170 			UVMHIST_LOG(maphist,"<- NOWAIT failed", 0,0,0,0);
171 			return(0);
172 		}
173 		simple_lock(&pager_map_wanted_lock);
174 		pager_map_wanted = TRUE;
175 		UVMHIST_LOG(maphist, "  SLEEPING on pager_map",0,0,0,0);
176 		UVM_UNLOCK_AND_WAIT(pager_map, &pager_map_wanted_lock, FALSE,
177 		    "pager_map", 0);
178 		goto ReStart;
179 	}
180 
181 enter:
182 	/* got it */
183 	for (cva = kva ; size != 0 ; size -= PAGE_SIZE, cva += PAGE_SIZE) {
184 		pp = *pps++;
185 		KASSERT(pp);
186 		KASSERT(pp->flags & PG_BUSY);
187 		pmap_kenter_pa(cva, VM_PAGE_TO_PHYS(pp), prot);
188 	}
189 	pmap_update(vm_map_pmap(pager_map));
190 
191 	UVMHIST_LOG(maphist, "<- done (KVA=0x%x)", kva,0,0,0);
192 	return(kva);
193 }
194 
195 /*
196  * uvm_pagermapout: remove pager_map mapping
197  *
198  * we remove our mappings by hand and then remove the mapping (waking
199  * up anyone wanting space).
200  */
201 
202 void
203 uvm_pagermapout(kva, npages)
204 	vaddr_t kva;
205 	int npages;
206 {
207 	vsize_t size = npages << PAGE_SHIFT;
208 	struct vm_map_entry *entries;
209 	UVMHIST_FUNC("uvm_pagermapout"); UVMHIST_CALLED(maphist);
210 
211 	UVMHIST_LOG(maphist, " (kva=0x%x, npages=%d)", kva, npages,0,0);
212 
213 	/*
214 	 * duplicate uvm_unmap, but add in pager_map_wanted handling.
215 	 */
216 
217 	pmap_kremove(kva, npages << PAGE_SHIFT);
218 	if (kva == emergva) {
219 		simple_lock(&pager_map_wanted_lock);
220 		emerginuse = FALSE;
221 		wakeup(&emergva);
222 		simple_unlock(&pager_map_wanted_lock);
223 		return;
224 	}
225 
226 	vm_map_lock(pager_map);
227 	uvm_unmap_remove(pager_map, kva, kva + size, &entries);
228 	simple_lock(&pager_map_wanted_lock);
229 	if (pager_map_wanted) {
230 		pager_map_wanted = FALSE;
231 		wakeup(pager_map);
232 	}
233 	simple_unlock(&pager_map_wanted_lock);
234 	vm_map_unlock(pager_map);
235 	if (entries)
236 		uvm_unmap_detach(entries, 0);
237 	pmap_update(pmap_kernel());
238 	UVMHIST_LOG(maphist,"<- done",0,0,0,0);
239 }
240 
241 /*
242  * interrupt-context iodone handler for nested i/o bufs.
243  *
244  * => must be at splbio().
245  */
246 
247 void
248 uvm_aio_biodone1(bp)
249 	struct buf *bp;
250 {
251 	struct buf *mbp = bp->b_private;
252 
253 	KASSERT(mbp != bp);
254 	if (bp->b_flags & B_ERROR) {
255 		mbp->b_flags |= B_ERROR;
256 		mbp->b_error = bp->b_error;
257 	}
258 	mbp->b_resid -= bp->b_bcount;
259 	pool_put(&bufpool, bp);
260 	if (mbp->b_resid == 0) {
261 		biodone(mbp);
262 	}
263 }
264 
265 /*
266  * interrupt-context iodone handler for single-buf i/os
267  * or the top-level buf of a nested-buf i/o.
268  *
269  * => must be at splbio().
270  */
271 
272 void
273 uvm_aio_biodone(bp)
274 	struct buf *bp;
275 {
276 	/* reset b_iodone for when this is a single-buf i/o. */
277 	bp->b_iodone = uvm_aio_aiodone;
278 
279 	simple_lock(&uvm.aiodoned_lock);	/* locks uvm.aio_done */
280 	TAILQ_INSERT_TAIL(&uvm.aio_done, bp, b_freelist);
281 	wakeup(&uvm.aiodoned);
282 	simple_unlock(&uvm.aiodoned_lock);
283 }
284 
285 /*
286  * uvm_aio_aiodone: do iodone processing for async i/os.
287  * this should be called in thread context, not interrupt context.
288  */
289 
290 void
291 uvm_aio_aiodone(bp)
292 	struct buf *bp;
293 {
294 	int npages = bp->b_bufsize >> PAGE_SHIFT;
295 	struct vm_page *pg, *pgs[npages];
296 	struct uvm_object *uobj;
297 	struct simplelock *slock;
298 	int s, i, error, swslot;
299 	boolean_t write, swap;
300 	UVMHIST_FUNC("uvm_aio_aiodone"); UVMHIST_CALLED(ubchist);
301 	UVMHIST_LOG(ubchist, "bp %p", bp, 0,0,0);
302 
303 	error = (bp->b_flags & B_ERROR) ? (bp->b_error ? bp->b_error : EIO) : 0;
304 	write = (bp->b_flags & B_READ) == 0;
305 	/* XXXUBC B_NOCACHE is for swap pager, should be done differently */
306 	if (write && !(bp->b_flags & B_NOCACHE) && bioops.io_pageiodone) {
307 		(*bioops.io_pageiodone)(bp);
308 	}
309 
310 	uobj = NULL;
311 	for (i = 0; i < npages; i++) {
312 		pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT));
313 		UVMHIST_LOG(ubchist, "pgs[%d] = %p", i, pgs[i],0,0);
314 	}
315 	uvm_pagermapout((vaddr_t)bp->b_data, npages);
316 
317 	swslot = 0;
318 	slock = NULL;
319 	pg = pgs[0];
320 	swap = (pg->uanon != NULL && pg->uobject == NULL) ||
321 		(pg->pqflags & PQ_AOBJ) != 0;
322 	if (!swap) {
323 		uobj = pg->uobject;
324 		slock = &uobj->vmobjlock;
325 		simple_lock(slock);
326 		uvm_lock_pageq();
327 	} else if (error) {
328 		if (pg->uobject != NULL) {
329 			swslot = uao_find_swslot(pg->uobject,
330 			    pg->offset >> PAGE_SHIFT);
331 		} else {
332 			swslot = pg->uanon->an_swslot;
333 		}
334 		KASSERT(swslot);
335 	}
336 	for (i = 0; i < npages; i++) {
337 		pg = pgs[i];
338 		KASSERT(swap || pg->uobject == uobj);
339 		UVMHIST_LOG(ubchist, "pg %p", pg, 0,0,0);
340 
341 		/*
342 		 * for swap i/os, lock each page's object (or anon)
343 		 * individually since each page may need a different lock.
344 		 */
345 
346 		if (swap) {
347 			if (pg->uobject != NULL) {
348 				slock = &pg->uobject->vmobjlock;
349 			} else {
350 				slock = &pg->uanon->an_lock;
351 			}
352 			simple_lock(slock);
353 			uvm_lock_pageq();
354 		}
355 
356 		/*
357 		 * process errors.  for reads, just mark the page to be freed.
358 		 * for writes, if the error was ENOMEM, we assume this was
359 		 * a transient failure so we mark the page dirty so that
360 		 * we'll try to write it again later.  for all other write
361 		 * errors, we assume the error is permanent, thus the data
362 		 * in the page is lost.  bummer.
363 		 */
364 
365 		if (error) {
366 			int slot;
367 			if (!write) {
368 				pg->flags |= PG_RELEASED;
369 				continue;
370 			} else if (error == ENOMEM) {
371 				if (pg->flags & PG_PAGEOUT) {
372 					pg->flags &= ~PG_PAGEOUT;
373 					uvmexp.paging--;
374 				}
375 				pg->flags &= ~PG_CLEAN;
376 				uvm_pageactivate(pg);
377 				slot = 0;
378 			} else
379 				slot = SWSLOT_BAD;
380 
381 			if (swap) {
382 				if (pg->uobject != NULL) {
383 					int oldslot;
384 					oldslot = uao_set_swslot(pg->uobject,
385 						pg->offset >> PAGE_SHIFT, slot);
386 					KASSERT(oldslot == swslot + i);
387 				} else {
388 					KASSERT(pg->uanon->an_swslot ==
389 						swslot + i);
390 					pg->uanon->an_swslot = slot;
391 				}
392 			}
393 		}
394 
395 		/*
396 		 * if the page is PG_FAKE, this must have been a read to
397 		 * initialize the page.  clear PG_FAKE and activate the page.
398 		 * we must also clear the pmap "modified" flag since it may
399 		 * still be set from the page's previous identity.
400 		 */
401 
402 		if (pg->flags & PG_FAKE) {
403 			KASSERT(!write);
404 			pg->flags &= ~PG_FAKE;
405 			uvm_pageactivate(pg);
406 			pmap_clear_modify(pg);
407 		}
408 
409 		/*
410 		 * do accounting for pagedaemon i/o and arrange to free
411 		 * the pages instead of just unbusying them.
412 		 */
413 
414 		if (pg->flags & PG_PAGEOUT) {
415 			pg->flags &= ~PG_PAGEOUT;
416 			uvmexp.paging--;
417 			uvmexp.pdfreed++;
418 			pg->flags |= PG_RELEASED;
419 		}
420 
421 		/*
422 		 * for swap pages, unlock everything for this page now.
423 		 */
424 
425 		if (swap) {
426 			if (pg->uobject == NULL && pg->uanon->an_ref == 0 &&
427 			    (pg->flags & PG_RELEASED) != 0) {
428 				uvm_unlock_pageq();
429 				uvm_anon_release(pg->uanon);
430 			} else {
431 				uvm_page_unbusy(&pg, 1);
432 				uvm_unlock_pageq();
433 				simple_unlock(slock);
434 			}
435 		}
436 	}
437 	if (!swap) {
438 		uvm_page_unbusy(pgs, npages);
439 		uvm_unlock_pageq();
440 		simple_unlock(slock);
441 	} else {
442 		KASSERT(write);
443 
444 		/* these pages are now only in swap. */
445 		simple_lock(&uvm.swap_data_lock);
446 		KASSERT(uvmexp.swpgonly + npages <= uvmexp.swpginuse);
447 		if (error != ENOMEM)
448 			uvmexp.swpgonly += npages;
449 		simple_unlock(&uvm.swap_data_lock);
450 		if (error) {
451 			if (error != ENOMEM)
452 				uvm_swap_markbad(swslot, npages);
453 			else
454 				uvm_swap_free(swslot, npages);
455 		}
456 		uvmexp.pdpending--;
457 	}
458 	s = splbio();
459 	if (write && (bp->b_flags & B_AGE) != 0) {
460 		vwakeup(bp);
461 	}
462 	pool_put(&bufpool, bp);
463 	splx(s);
464 }
465